Closes#24514.
The original PR was _supposed_ to be handling this, and had code that
_looked_ like it would handle it, but was in fact subtly wrong, as the
union of `RectangleF.Empty` (initial value of `area`) and an arbitrary
other rectangle is _not_ the second rectangle, but rather a rectangle
that contains _both_ the second rectangle and the point (0,0),
thus leading to the bound calculation fail.
I don't see an issue with applying this workaround to more places, even
though it is a workaround, because it marks each usage very clearly. If
we design a better solution in the future it should be easy to replace
the usages.